Custom models. Delivered, not rented.
Every service ends the same way: you receive a model you own, evidence it works, and a recipe to run it where you want. Hosting is your choice; we help with the first deploy.
Custom fine-tuning
We adapt an open-weight model to your task using your examples. In plain words, we show the model many pairs of "this input, this correct output" until it does the job your way. Three techniques, chosen by the data you have:
| You have | We use | Good for |
|---|---|---|
| Examples of the right output | Supervised fine-tuning (SFT) | Extraction, classification, formats, tool calls, house style. The default. |
| Pairs of "this answer is better than that one" | Preference tuning (DPO / ORPO) | Tone and quality when you can rank answers but cannot write the one true answer. |
| A way to check if an answer is correct | Reinforcement learning (GRPO) | Maths, code, agents and other verifiable tasks with few labels. |
Base models are open-weight and licence-clean (Qwen family by default, Apache 2.0). Sizes from 0.6B for narrow structured tasks to 9B for writing and reasoning, plus an 8B vision model for documents and images.
Personal models
An assistant that writes, sorts and replies the way you do, trained on your own mail, notes and documents. Delivered as a small quantized model that runs on your laptop with Ollama or a similar tool, with no cloud account needed.
- Typical size: 0.6B to 4B parameters, runs on a normal laptop.
- Delivery: one file plus a one-line install.
- Optional monthly retrain from the corrections you make.
Workflow specialists
One trained model per workflow: quotes, intake forms, support triage, invoice fields, catalogue clean-up. Each specialist is scored on your own held-out examples before it ships, and runs on one small GPU or a managed service like AWS Bedrock.
- Start with the workflow that costs the most hours.
- Before/after numbers on every delivery.
- Add specialists over time; they share the same base and the same deploy.
Model fleets
Many departments, tenants or languages, each with its own small adapter, all served from one base model on your private deployment. Each adapter is a few megabytes, so a single GPU can hold dozens of specialists and pick the right one per request.
- Training manifest for every adapter: base model, data hash, settings, who approved it.
- Eval pack and failure examples for your review process.
- Private hosting on AWS, Cloudera or on-prem; optional on-site training.
Deploy anywhere
You receive the same bundle whatever the target: merged weights in standard Hugging Face format, the adapter on its own, a quantized copy, the eval report, a deploy recipe and the training manifest.
| Target | What we hand over |
|---|---|
| AWS Bedrock (Custom Model Import) | Merged weights in the supported architecture and library versions; import checklist. |
| Amazon SageMaker / EKS | vLLM container recipe, adapter stacking for fleets. |
| Cloudera AI Inference | Hugging Face checkpoint for the vLLM route or NIM packaging. |
| On-prem | vLLM or NIM recipe, GPU sizing, health check. |
| Laptop | GGUF for Ollama or llama.cpp. |
Retrain subscription
Work changes. With your permission, corrected examples flow back through a small hook in the deploy recipe, and we retrain on a schedule you set. Same manifest, new data, new eval report. You stay in control of what is sent.